Test-Time Scaling Selection on RouterBench Cost-Sensitive
0.8337RewardOracle
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| OracleNumber of seeds=5, Iterations=160, Warm-up phase=50-step2026.05 | 0.8337 | 57.38 | 10.6 | |
| OracleBackbone=Qwen3-0.6B, Number of seeds=5, Iterations=160, Warm-up phase=50-step2026.05 | 0.8297 | 54.87 | 10.5 | |
| UNISCALEBackbone=Qwen3-0.6B, Number of seeds=5, Iterations=160, Warm-up phase=50-step2026.05 | 0.7535 | 46.5 | 23.3 | |
| GreedyBackbone=Qwen3-0.6B, Number of seeds=5, Iterations=160, Warm-up phase=50-step2026.05 | 0.7184 | 43.75 | 54 | |
| UNISCALENumber of seeds=5, Iterations=160, Warm-up phase=50-step2026.05 | 0.7079 | 46.88 | 49.4 | |
| TSBackbone=Qwen3-0.6B, Number of seeds=5, Iterations=160, Warm-up phase=50-step2026.05 | 0.7022 | 44.5 | 67.9 | |
| MLPBackbone=Qwen3-0.6B, Number of seeds=5, Iterations=160, Warm-up phase=50-step2026.05 | 0.7006 | 41.75 | 48.6 | |
| NeuralUCBBackbone=Qwen3-0.6B, Number of seeds=5, Iterations=160, Warm-up phase=50-step2026.05 | 0.6984 | 42 | 70.6 | |
| k-NNBackbone=Qwen3-0.6B, Number of seeds=5, Iterations=160, Warm-up phase=50-step2026.05 | 0.6966 | 36.5 | 49 | |
| RandomBackbone=Qwen3-0.6B, Number of seeds=5, Iterations=160, Warm-up phase=50-step2026.05 | 0.6937 | 42.12 | 76.7 | |
| k-NNNumber of seeds=5, Iterations=160, Warm-up phase=50-step2026.05 | 0.659 | 41.38 | 326 | |
| MLPNumber of seeds=5, Iterations=160, Warm-up phase=50-step2026.05 | 0.6301 | 47 | 644 | |
| NeuralUCBNumber of seeds=5, Iterations=160, Warm-up phase=50-step2026.05 | 0.588 | 50.75 | 1,558.2 | |
| RandomNumber of seeds=5, Iterations=160, Warm-up phase=50-step2026.05 | 0.5731 | 53 | 1,358.7 | |
| GreedyNumber of seeds=5, Iterations=160, Warm-up phase=50-step2026.05 | 0.5589 | 45 | 660.6 | |
| TSNumber of seeds=5, Iterations=160, Warm-up phase=50-step2026.05 | 0.5549 | 52.12 | 1,762.8 |